Forecasting Groundwater Quality Parameters using Machine Learning Models: a Case Study of Khemismiliana Plain, Algeria

Abstract

Total Dissolved Solids, one of the most extensively used indicators for assessing groundwater quality, it useful to estimate salinity and hardness in water. The objective of the present study is to develop accurate and dependable machine learning models for forecasting the total dissolved solids, parameter; as well to evaluate and explain the relationship of total dissolved solids with the mineral salts. Four machine learning models Decision tree, Random forest, Adaboost and support victor regression SVR have been successfully employed for modeling the total dissolved solids using Electrical Conductivity (EC) and concentrations of major elements (Ca²⁺, Mg²⁺, Na⁺, K⁺, Cl^–, SO$_{4}^{{2 - }}$, HCO$_{3}^{ - }$, NO$_{3}^{ - }$) of the groundwater aquifer in upper Cheliff plain (the northwestern of Algeria). One hundred ninety-one of observations collected from wells by the ANRH (national water resources agency, Algeria) for a period of 8 years between 2008 and 2016, were randomly divided into training and validation sets. The overall prediction performance results indicated that the models provided satisfactory estimation with priority to the support vector regression model, based on the four parameters including: EC, Na⁺, SO$_{4}^{{2 - }}$, and Cl^–, with the best support vector machine results of RMSE = 0.0328; NS = 0.9455. Feature selection method revealed that the correlation analysis results were reliable and could be utilized as a first step in selecting the optimum input data for forecasting groundwater quality parameters. Generally, the proposed models are useful in predicting groundwater quality parameters and may aid decision-makers in developing and managing groundwater plans.